Author: Claude, DeepChain TechFlow
DeepChain Summary: On August 12, OpenAI and five scholars released a 69-page working paper linking over 17 million real usage logs from ChatGPT Enterprise with employee seniority levels, task categories, and financial data of publicly traded companies. The findings contradict conventional wisdom: data refutes the narrative that “AI will first replace junior employees,” revealing that young employees with just a few years of experience are the most intensive users across all levels. Meanwhile, AI adoption is widening—rather than narrowing—corporate disparities, with adopters’ median income, market capitalization, and R&D spending all ten times higher than those of non-adopters.
On August 12, a working paper titled "How Organizations Use AI: Evidence from ChatGPT" was posted on arXiv. The authors—Aaron Chatterji from Duke University, Prasanna Tambe from the Wharton School, David Holtz from OpenAI, and three other scholars—produced a 69-page study. Unlike typical industry surveys, this paper links ChatGPT Enterprise account logs with employee seniority levels, message content categories, and financial data from U.S. public companies, analyzing over 17 million messages across more than 1,500 organizations within a six-month observation window. In other words, OpenAI has, for the first time, directly shared its enterprise customers’ real usage data with the academic community and the public.
The most active users aren't executives—recent hires send eight to nine more messages per week than the company average.
By headcount, early-career employees and trainees make up only 7% of weekly active users—the smallest group; managers and directors account for 24%, and executives for 10%. But when shifting the focus from “who is using” to “who is using most intensively,” the picture reverses entirely: early-career employees send 8 to 9 more messages per week on average than their company’s typical user—the highest intensity of any level—while executives, founders, and partners send fewer messages than the company average.
This means that the popular concern that “AI will first replace entry-level workers” has not been observed within large corporations. Those assumed to be most vulnerable are precisely the ones who are using AI as their primary tool. The paper’s authors specifically address this contrast: entry-level employees are closest to the grunt work, have the least entrenched work habits, and experience the greatest marginal gains from tools—making them the fastest learners and most aggressive adopters.
True growth lies not in acquiring new customers, but in deepening engagement with existing ones.
The total figures are even more striking. Between June 2025 and March 2026, total token consumption for ChatGPT Enterprise increased by approximately sevenfold. Breaking it down, among companies that adopted the service before June 2025, internal token usage also rose by about fourfold—indicating that a significant portion of the growth stems from existing customers deepening their usage, rather than new customer acquisition.
More notably, in early 2026, enterprise adoption accelerated simultaneously across different time points. This does not resemble individual companies gradually discovering the product on their own; rather, it appears as if the product’s capabilities as a whole took a leap forward, lifting usage intensity for everyone at once. For enterprises, the implication of this data is clear: purchasing ChatGPT is just the entry ticket—the real variable is whether usage deepens over time.
A tenfold gap: Early entrants are pulling far ahead of their competitors.
Zooming out to the corporate level, the most striking part of the report emerges. Among U.S. public companies, those using ChatGPT Enterprise show a tenfold advantage over those that don’t: median revenue of $2.275 billion versus $210 million, median market capitalization of $5 billion versus $316 million, and median R&D spending of $113 million versus $9.9 million.
Companies in the top 5% by revenue are 9.8 percentage points more likely to adopt AI; even when comparing within the same industry, the top 5% are 11.3 percentage points more likely. Moreover, the companies that adopt AI most deeply are precisely those with the highest market value per employee, and this holds true even after adjusting for industry and size factors. The paper’s own conclusion is straightforward: during the early stages of AI diffusion, existing disparities between firms will be amplified, not reduced.
For investors, this report offers a rare real-world example: the market is already rewarding companies that deeply integrate AI—there’s a visible gap in financial metrics between those that bought thousands of licenses and let them gather dust as mere “AI concepts,” and those that have truly embedded AI into their workflows.
There is no killer app—just a long tail of 60 tasks.
So what do businesses actually use AI for? The paper categorized over 17 million messages into 60 types of tasks: more than half of active users use it weekly for documentation and technical writing, nearly half use it for technical and digital work, but the remaining uses are highly diverse—with large numbers of users also employing it for communication, topic research, fact-checking, sales and marketing, planning, legal tasks, data analysis, and financial and tax matters.
Tasks and roles are strictly aligned: engineers primarily handle technical work and debugging, finance staff manage financial and tax tasks, and sales and marketing teams focus on sales and marketing—but no role-specific task overshadows the common tasks used by everyone. Industry differences also exist: financial and insurance industries have a significantly higher proportion of financial and tax tasks, while retail, information, and entertainment industries concentrate more on sales and marketing tasks. The same tool evolves differently across industries.
This also explains why no "killer app"-style workflow has emerged. In enterprises, AI takes the form of a backing layer beneath all knowledge work—each role takes a small piece, and only together do they create real penetration.
AI hasn't closed the gap—it has split companies into two worlds: those that know how to use it and those that don't.
The most striking line in the report is hidden in the conclusion: In the early stages of AI adoption, it will widen, not narrow, the gap between companies. Initially, many believed AI was a tool for leveling the playing field, but data shows it acts more like a magnifying glass—those who get on board early and use it deeply pull further ahead. For individuals, data has debunked the anxiety that “AI will first replace entry-level workers,” but a new risk is just emerging: AI won’t replace everyone—it will replace those who refuse to adopt the tool. Young people have already written the answer in their message logs; now it’s up to everyone else to catch up.